What the documents look like
Brand outlet discount operations exist to clear out-of-season inventory. The working materials are discount approval forms, slow-moving inventory lists, price tags and gross-margin worksheets, and they refresh on a weekly and monthly cadence. Copy drafting draws on these materials for discount ranges, inventory-age status and margin definitions, so the retrieval step has to return the original wording for a given subject, object and validity scope.
The materials are mainly Chinese tables and approval records. Style numbers, item codes and discount tiers are exact-query objects. Natural-language phrasing is used for describing why stock is slow-moving and for clearance scenarios. Retrieved content serves as drafting evidence only. It does not decide approval, authorization or whether a price takes effect.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The vector model identifier must be checked first and non-matching bases excluded, with the actual observation recorded. A cross-model search is not run to compare missed recalls.
Search modes are semantic search, full-text search and hybrid search. Hybrid search merges multiple result paths through RRF. Reranking requires a model and must be enabled in settings. Image retrieval requires image input and the corresponding model capability. The optional paths shown in the document flow do not mean every request executes all of them. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Cited content is output as an array and its length can be zero, so the empty-result path still completes. Retrieval and similarity scores do not constitute permission or entity isolation. Business verification, approval and writes must be implemented downstream.
Hypothetical paper exercise: an outlet store discount approval form states "Style A-231, inventory age 210 days, approved discount range 30% to 40% off, approval validity limited to this store's August 2024 clearance event." The accompanying slow-moving inventory list states "Style A-231, status is pending review." A draft sentence reads "Style A-231 has been approved for clearance at 30% off and can be published externally." Checking requires comparing the subject, object, status and validity scope in the approval form. Pending review does not equal approval. An approved range does not equal an executed discount. This sentence is not supported by the materials.
Configuration and implementation checks
| Configuration or check item | Use on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases with the same vector model | Discount approval and slow-moving list must be associated under one vector model |
| Search mode | Compare semantic, full-text and hybrid search against the test question | Style A-231 is an exact query, so full-text or hybrid search is used for checking |
| Citation limit | Unit is token; test completeness with actual materials of different lengths | Starting point for this hypothetical exercise is 2000, because the approval form and list are short in total; compare against 1000 to see whether content stays complete |
| Minimum relevance | Takes effect only in semantic search or when result reranking is used | Mark relevant and irrelevant samples first, then adjust the threshold; it is not used to exclude unrelated businesses |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | After a model is available and the setting is enabled, candidates are reordered | Enable when candidates are numerous; the score is used for filtering and does not replace manual checking |
Three common errors
- Treating pending review as approved. The input is the slow-moving list entry "status is pending review." The comparison object is the approval record in the approval form. If the list has no approval field, the claim is unsupported and no publishable conclusion may be output.
- Treating the approved discount range as the executed discount. The input is "approved discount range 30% to 40% off." The comparison object is "at 30% off" in the draft sentence. A range contains multiple values. If no executed value is stated, the item remains to be verified.
- Using a similarity setting to exclude unrelated businesses. The input is a minimum relevance number. The comparison object is the retrieval scope. This value only filters low-relevance results and does not constitute permission or entity isolation. Unauthorized records may still be recalled.
How to verify the setup
- Input the approval form text from the hypothetical sample and the draft sentence. Compare whether subject, object, status and validity scope match. If the draft sentence says "has been approved" while the list states only "pending review," the claim is unsupported and the evidence gap must be recorded.
- Input style number A-231 for full-text or hybrid search. Compare whether the returned citations contain the approval record for that style number. If the citation is an empty array, check the downstream design against the empty-result path. Do not infer that a rule-based search necessarily returns empty.
- Input two knowledge bases carrying the same vector model identifier for an association check. Compare whether the identifiers match. If they do not match, exclude that base and record it. Do not run a cross-model search and do not infer missed recalls or redundant results.
The public technical facts checked for this feature are: associated knowledge bases must share one vector model; search modes are compared against the test question without assuming one mode suits all materials; the citation limit is measured in tokens and is tested with actual materials of different lengths; minimum relevance takes effect only in semantic search or with reranking, and thresholds are set after marking relevant and irrelevant samples; question optimization handles coreference resolution and question expansion, and the original and optimized questions are compared to check whether the subject was altered.